OpenAI Launches GPT-6 Astra: Enhanced Security and Long-Context Features

Key Takeaways

1GPT-6 Astra is being deployed today for a limited number of organizations

2High scores announced in security and long context processing

3Artificial Analysis ranks it behind Claude Fable 5.1 in Intelligence, but ahead in cost-effectiveness for coding

Why it matters : GPT-6 Astra aims to compete with Fable by offering advanced performance in security and long context, while also presenting aligned pricing and expanded availability.

OpenAI begins the rollout of GPT-6 Astra, initially for a limited number of organizations, and soon for ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as via the API and AWS. API pricing is set at $10/million input tokens and $50/million output tokens, with the label "gpt-6-astra." Astra scores high in security and on very long contexts, but still trails behind Claude Fable 5.1 on the Intelligence Index from Artificial Analysis, while being well-positioned in cost-effectiveness for code agents. On ARC-AGI 3, a score of 99.9% was achieved with a proprietary adapter for $19,000.

Artificial Analysis credits Astra with 61 in Intelligence and a cost-effectiveness advantage in code

Artificial Analysis assigns GPT-6 Astra a score of 61 in its Intelligence Index, tying with GPT-5.6 Sol and five points below Claude Fable 5.1, which reaches the maximum with feedback in this ranking. Astra is also ranked behind Meta's Muse Spark 1.3 in this index. On the Coding Agent Index, Artificial Analysis places Astra at the forefront of cost-effectiveness: at maximum effort, the model costs about the same as GPT-5.6 Sol while achieving two points more, and for each task, it costs less than half the price of Claude Fable 5 for an equivalent score.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

OpenAI announces high scores in security and on very long contexts

OpenAI reports that GPT-6 Astra achieves 100% on ExploitBench, compared to 78.5% for GPT-5.6 Sol. On ExploitGym, Astra reaches 42.4%, while Sol scores 30.3%. In binary reverse engineering on SRE-Bench, Astra shows 99.2% over four attempts, compared to 68.7% for Sol. For processing very long contexts, OpenAI indicates 100% between 256K and 512K tokens and 96.3% between 512K and 1M tokens on its "eight needles" benchmark. A recent incident involving Hugging Face is mentioned in the context of security performance, without further details.

ARC-AGI 3: 99.9%, but discrepancies based on adapter and run cost

OpenAI states that GPT-6 Astra achieves 99.9% on ARC-AGI 3, a benchmark published in March, while Claude Fable 5 has not yet published results on this test. According to the ARC-AGI blog, achieving this 99.9% score required $19,000 with OpenAI's "Provider Adapter harness," while the standard ARC-AGI adapter achieved 62.7% for a cost of $26,000. OpenAI indicates that the Provider Adapter maintains a non-transparent reasoning state between queries and applies compaction to allow for the reuse of work during extended discussions.

Gradual availability and announced pricing on API and AWS

The rollout of GPT-6 Astra begins today for a limited number of organizations. Access will be expanded in the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as via OpenAI's API and AWS. API pricing is set at $10 per million input tokens and $50 per million output tokens. OpenAI claims that GPT-6 Astra outperforms Fable on most of its self-reported benchmarks. The model label intended for the API is "gpt-6-astra."